SDE-YOLO: An Improved YOLOv8 Algorithm for Weed Detection in Crop Fields

Sesame is an important oilseed crop, but weed infestation can adversely affect its growth and yield. Missed weed detections may compromise subsequent control operations, making reliable and comprehensive weed detection essential for precision agriculture. To reduce missed detections while maintaining low model complexity, this study proposes SDE-YOLO, a lightweight detector based on YOLOv8n. First, C2f_Star modules are introduced into the backbone to enhance feature representation while reducing model complexity. Second, DASI modules are incorporated into the neck to improve multiscale feature fusion and strengthen the representation of targets under complex field conditions. Finally, EIoU is adopted for bounding-box regression to provide more effective geometric optimization. Experiments on the Crop and Weed Detection Data with Bounding Boxes dataset show that SDE-YOLO achieves a recall of 78.60% and an mAP@0.5 of 84.40%, outperforming YOLOv8n by 6.00 and 2.40 percentage points, respectively. Meanwhile, the parameter count and GFLOPs are reduced by 18.60% and 3.70%, resulting in only 2.45 M parameters and 7.80 GFLOPs. Additionally, generalization experiments were conducted on two independent datasets, UAV-PDD2023 and WeedCrop Image, to further evaluate the applicability of the proposed SDE-YOLO model beyond the primary dataset. These results indicate that SDE-YOLO improves detection completeness and overall detection performance while retaining a compact model structure, making it a promising solution for precision weed management.